AEGIS · Autonomous Executive Guild Intelligence System

Self-hosted AI agents that ask before they act

AI systems built around the way you work, not the other way round

AEGIS, the Autonomous Executive Guild Intelligence System, is the set of self-hosted AI agents I run my own business on: bookkeeping, infrastructure, mail, research, publishing and analytics. Each one is shaped around how I actually work, and each keeps a human in the loop: it asks me before it changes anything. It is also the kit I use to build the same kind of system around the way you work.

Check your workflow firstSee what it runs

The problem

Most AI tools ask you to change how you work

A tool is built for its median customer. If your work looks like theirs, buy it. If it does not, you pay for the whole tool, use part of it, and change your process to fit the rest: export here, re-key there, a spreadsheet to bridge the gap, a person to check what the tool cannot.

That gap is where my own work sat. None of the tools fitted an Indian bank feed, a Docker Swarm instead of Kubernetes, or a plain-text ledger. So I built the systems around the work instead.

The AI bookkeeping tools assume QuickBooks or Xero and a US bank feed.

Books reads Indian bank SMS alerts, e-statement mail and password-locked PDFs, and writes a plain-text hledger journal.

The AI SRE tools assume Kubernetes.

DevOps reads Prometheus and Alertmanager on a small Docker Swarm, and acts only through ssh keys locked to one script each.

The inbox assistants want your whole mailbox in their cloud.

Assistant triages Gmail, IMAP and Outlook from my own server and writes to the Todoist projects I already use.

Approval steps sit in the enterprise tier, if they exist at all.

Every write that leaves the system waits for my answer on a card, in every system, from day one.

What it runs

Seven jobs, each its own system

Each system does one job end to end, with its own database and its own app. They talk to each other over small, authenticated routes, and none can touch another’s data. Every number below comes from running them on my own work since October 2026.

Bookkeeping for an Indian founder

Books

Bank statements, alert mail and SMS in; reconciled entries in an hledger journal out. A payee it does not know becomes one question on a card, and the answer becomes a rule, so it asks less each week.

Matched the old system on all 15 statements before it was allowed to write. Along the way it showed the old system had missed about 44 UPI alerts a week.

Alert to fix on a homelab

DevOps

Takes every alert, groups duplicates, investigates read-only, and proposes a restart, scale or rollback on a card. One careful automatic restart is the only thing it may do alone.

Replayed the old system’s alerts before cutover: 143 of 148 problems came out the same, and each difference was explained. Its model budget is capped at $20 a month.

Mail, calendar, tasks and meetings

Assistant

Triages three kinds of mailbox, keeps my task list, preps meetings from what it knows about the people in them, and sends a brief at 08:00.

Dry-ran an archive of 7,417 old mails first, then archived 746 and kept 148. Its rule: informational mail may go, important mail never.

A library that answers with sources

Research

Feeds, bookmarks, papers and my book library go into one search. It sends a daily digest and answers questions with citations.

About 37,000 documents imported from the old system.

Issue to draft pull request

Development

Takes an issue from me or from another system, plans it, and runs a coding agent in its own git worktree. It asks before it starts and again before it opens a draft PR. Merges stay mine.

The other systems report faults in their own code to it, so the family files its own bugs.

Search data in, posts out

Product and Social

Product reads GA4, Search Console and Bing for about two dozen sites, queues pages for indexing and drafts issues each week under a cap. Social syndicates posts with a canonical link back, and holds the full copies until Google has indexed the original.

Every post passes an approval card. It never comments on anyone’s post for me.

One place to answer them

Mission Control puts every system’s cards in one inbox, so I answer a journal question, a restart and a post in the same place. I can talk to any one system, or hold a meeting with all of them and get minutes at the end. It runs in the browser and as a desktop and phone app built from the same code, with voice in both directions. It is for me only: the same login check guards every system’s own API, not just the edge.

Mission Control on the phone: cards that need me, raised hands, apps and a meeting.

Watch it work

Self-hosted AI agents at work, two minutes each

Short demos recorded in Mission Control. The data on screen is invented for the demos; the rules, the flows and the mistake are the real ones.

Alert to pull request, four tapsA crash-looping worker at 2 a.m. becomes a draft pull request by morning, after four taps from me.
The agent said noFour times the systems refuse, and why that is the whole point.
A question becomes a ruleBookkeeping asks about a payee once. The answer becomes a rule and re-files every past entry it matches.
The assistant’s first mistakeThe inbox sweep got it wrong the first time. What changed, and why every switch now runs dry first.
Social waits for GoogleSocial will not republish a post until Google has indexed the original, and leaves the LinkedIn draft for me to write.

The kit

Why a custom AI agent no longer takes months

Every system shares one SDK and one UI kit: login, credentials, cards, schedules, model budgets, audit lines and the connectors to common tools are built and tested once. A new system is mostly the part that is specific to the work.

Your tools, by configuration

A system asks for “mail for bills” or “files for statements”, not for Gmail or Drive. Which product answers is one row of configuration, so the same code reads Gmail for me and a forwarded IMAP mailbox for someone else.

Approval on every outside write

A model can only propose. A journal entry, a restart, a post or a pull request is a typed action with a dry run, a card and a state. An action that fails in an unclear way is never retried on its own.

Shadow first, then switch over

A new system runs beside the old process, writing nowhere that matters, until its output matches line for line. The old process is switched off before the new one writes, so there is never a double entry.

Known model costs

Every model call goes through one module that checks a daily budget before the call and records it after. You choose the models, local ones included, and each deployment has its own capped key.

Your deployment, your data

Each system runs as its own deployment with its own database, workflow namespace, encrypted credentials and search index, on infrastructure you control. Nothing is shared with another customer.

Durable by default

Schedules and long jobs run on Temporal, so a crash or a deploy does not lose work halfway. You can pause, resume or run any schedule from the app.

From first commit to live

How long each system took from its first commit to doing real work, in October 2026:

2 days
Books
2 days
DevOps
1–2 days
Assistant
up to 2 days
Research
same day
Development
2 days
Social
3 days
Product
same day; phone app 6 days
Mission Control

To be straight about these: they are my own systems, rebuilt by one engineer with coding agents, with the old version as a working reference to check every output against. A system for your business starts from your workflow, not from a working copy, so plan in weeks, not days.

Build or buy

When you should not hire me for this

A custom system is not always the answer. Often the honest one is in the middle: buy the standard layer, and build only the thin part that fits your work on top. AEGIS does exactly that. It posts through Postiz, keeps books in hledger and reads Prometheus; it does not rebuild them.

Build around your workflow when

  • The work is how you make money, not a back-office standard.
  • You pay for two or more tools that each do part of the job, and people glue the rest by hand.
  • The data should not leave your infrastructure.
  • Your stack is off the median: local banks, plain-text files, self-hosted tools.
  • The actions carry real risk, so you want to approve each one, on your terms.
  • Someone will own it after launch: your engineer, or me on a retainer.

Buy a product when

  • The process is standard and the vendor’s way is fine.
  • You need a compliance report such as SOC 2 now.
  • You need dozens of integrations you will never build.
  • Nobody will own the system once it is live.
  • You need it working this week.
  • A large team needs single sign-on and admin roles out of the box.

Where it started

Version 1 was open source. Version 2 is private.

Version 1 was one system doing everything for one person: one database, one worker, one deploy. I open-sourced it under the MIT licence so anyone could read how the approval gates work. I shut it down on 8 October 2026, once its jobs had moved to version 2, and archived the repo. It is still public and readable, but it gets no fixes, so treat it as a reference, not something to run.

Version 2 split each job into its own system on a shared kit, so one system can be built for one business without carrying the rest. Its code is private.

Read the archived v1 code →How v1 worked →

The build log

Notes from building it, newest first.

AEGIS v1 is archived. Here is what replaces it

I'm archiving the open-source AEGIS repo. Version 2 splits one system that did everything into separate products on a shared SDK, each with its own database and deploy. Why, what is live, and what happens to the old code.

Why AEGIS moved off my local GPU — and still runs on yours

I ran my personal AI system's thinking on an 8GB GPU at home for months, then moved it to Bedrock. Not because local models are bad — because my GPU is small and the work is bursty. Here are the production numbers, and why the embeddings stayed home.

LLM agents in production: a teardown of AEGIS, my open-source automation layer

Three LLM agents, 49 Temporal workflows, every risky action behind a human approval gate. An architecture teardown of AEGIS v1, now open source and archived — and what it teaches about production AI automation.

AEGIS Is Open Source

Three months ago I wrote about my personal AI orchestration system and ended with a section titled 'Why It's Not Open Source Yet.' That's fixed. AEGIS is on GitHub, MIT-licensed — here's what it actually took to get there.

Behavior Is Data, Not Code

The biggest refactor of the AEGIS open-sourcing sprint: removing every branch on an agent's identity so behavior lives in the database as capability tags — resolved at runtime, edited from a UI, and safe even when a tag has no owner.

A To-Do Is a Tweet: Social Publishing With Approval Cards

Scheduling a social post doesn't need a new UI. A to-do already has copy, a time, and labels — so in AEGIS, a Todoist task with a publish label is a scheduled post, and nothing goes out until a card in chat gets a human tap.

Bring Your Own Cloud: The Infrastructure Registry

For a self-hosted system that's meant to be forked, there's exactly one honest way to handle infrastructure credentials: the user brings their own, the system stores them encrypted, and nothing in the code assumes a vendor. Here's how AEGIS got there — and what it cost me to cut my own vendors out.

Labels, Not Projects: Rethinking GTD

The day Next and Someday stopped being Todoist projects and became labels — and why that small modelling decision is what made AEGIS's GTD layer, and its bidirectional sync, actually work.

Day One of Opening the Box

AEGIS is going open source, and today the first public commits landed. Day one of turning a private system into a shippable one: twenty-seven migrations squashed into a baseline, credentials evicted from the build, and an admin panel redesigned around decisions.

When the Bot Learns to Stay Quiet

AEGIS watches my homelab and GitHub, but the interesting part is what happens between an alert firing and me hearing about it. Most of the time, the correct answer is nothing.

One Primitive for Every Interruption

The design decision in AEGIS I'm proudest of isn't an agent or a model — it's a single interactions primitive. How one Postgres table, five kinds of card, and a Temporal workflow replaced every per-domain approval pattern.

Meet AEGIS: My Weird Little Operating Layer

AEGIS is my personal operating layer: four agents that watch my inbox, repos, alerts and feeds, and bring a proposed next step instead of a notification.

Building AEGIS: How Prolog + a Knowledge Graph Make Local LLMs Actually Useful

Prolog and a knowledge graph take routing, classification and fact lookup off the LLM, so most of a personal AI system runs on a cheap local model.

Questions

What people ask about AEGIS

Is AEGIS open source?
Version 1 is: MIT-licensed at github.com/hikmahtech/aegis, archived in October 2026, so it is public and readable but gets no fixes, releases or support. Version 2, the family of systems that runs today, is private.
Can I buy AEGIS or sign up for it?
No. There is no hosted version, no plan and no sign-up. Each system runs as one deployment on infrastructure its owner controls, with its own database, credentials and model keys. What I sell is building a system like it around your own workflow, as a consulting engagement; the scope and price are on the AI automation engineer page.
Does AEGIS run on self-hosted or local models?
Yes. Every model call goes through one proxy that maps tiers (fast, balanced, smart) to models, so routine work can run on local models and only heavier jobs escalate to a hosted one. You choose the models, each deployment has its own capped key, and embeddings run on my own hardware.
How does an AEGIS system ask for approval?
With a card: one question, a few buttons, in chat and in Mission Control. The workflow that raised it waits durably on Temporal, so the answer can come hours or days later and nothing is lost on a restart or a deploy. An answer can become a rule, so the same question is not asked twice.
What is human-in-the-loop AI, and how does AEGIS do it?
Human-in-the-loop AI means a person decides at the points where a mistake would cost something, and the model does the rest. In AEGIS that point is every write that leaves a system: a journal entry, a restart, a post or a pull request. The model investigates and proposes; a card asks one question; the action runs only after the answer. Routine reads and drafts need no approval, so the person is asked a few times a day, not a few hundred.
What does an agentic workflow look like in production?
Less like a chatbot and more like a colleague with a narrow job. An alert arrives, the agent groups duplicates, reads logs and metrics without changing anything, and proposes a restart on a card with its reasons; a tap runs it and a second check confirms the service is back. A bank statement arrives, the agent matches each line to the ledger, and the one payee it does not know becomes a question whose answer becomes a rule. The systems on this page are the examples, each with the numbers from running it.
How much does it cost to build an AI agent like this?
Two costs, kept apart. Building: my own systems took one to three days each from first commit to live work, because the kit (login, credentials, cards, schedules, model budgets, connectors) was already built and tested; a system around someone else’s workflow starts from that workflow, so plan in weeks, and the scope and price of an engagement are on the AI automation engineer page. Running: every model call is budgeted per day and recorded, so a system has a known monthly ceiling; the infrastructure agent runs under a cap of $20 a month in model calls, and routine work can run on local models for nothing.
Is AEGIS an alternative to OpenClaw or n8n?
Not as something you install. OpenClaw is an open-source personal assistant you run on your own machine and extend with skills; n8n is a workflow builder. Both are products with a community behind them, and if the work is standard and you want something running tonight, one of them is the right call. AEGIS is a way of building: separate self-hosted systems, one per job, each with its own database and an approval card on every outside write, built around a workflow that no tool fits. The build-or-buy section on this page says when each makes sense.
What happens when an AEGIS system gets something wrong?
Every write that leaves a system is typed: it is dry-run first when new, recorded with a state, and never retried on its own after an unclear failure, so it can be found and reversed. The inbox sweep that archived 2,466 mails on its first run is the worked example in the demos: the fix was a keep list and a cheaper second check, not a better prompt, and every archived mail can be brought back with one search.
What does AEGIS run on?
Python, FastAPI and Postgres, Temporal for durable workflows and schedules, Pydantic AI for agent turns inside workflow activities, a LiteLLM proxy for every model call, and React for each app. It is deployed on a small Docker Swarm behind Cloudflare Access, one deployment per customer, with no cloud region and no SaaS in the loop.
Is this the same AEGIS as the firewall, the naval system or the other AI products called Aegis?
No. AEGIS here stands for Autonomous Executive Guild Intelligence System, the name I gave the AI systems I run my own business on, first written up in April 2026. It is unrelated to the Aegis combat system, Aegis-branded security or firewall products, and the AI startups that use the name.

Have a system built around your work

Start by checking the workflow yourself. If it still looks worth building, tell me how the work runs today, and I will reply in writing with what I would build and what I would buy instead.

  1. 1Run the 17-question AI workflow teardown. Free, no email needed.
  2. 2Describe your workflow by email: the steps, the tools, what goes wrong. I answer each one myself.
  3. 3Then, if it fits, book a call.

What an engagement looks like, scope and price: AI workflows built and run on your own data →